Agent-System-Backend/internal/runtime/e4_escalation_test.go

package runtime

import (
	"context"
	"encoding/json"
	"strings"
	"testing"

	core "git-ssh.u1680832.nyat.app/weichen/Agent-System-Core.git"
)

// ---- E4：R4 升级提醒 ----

// 滑动窗口判据：近 6 次结果中 3 次失败（可与成功交错）触发换路指令；
// 每 run 一次；2 次失败不触发。
func TestEscalationOnConsecutiveToolFailures(t *testing.T) {
	st := &directRun{l: &LLMRunner{}, inflight: map[string]*toolSlot{}}
	st.noteFailureOutcome("ok")
	st.noteFailureOutcome("error")
	st.noteFailureOutcome("ok")
	st.noteFailureOutcome("error")
	if got := st.escalationReminders(); len(got) != 0 {
		t.Fatalf("two scattered failures must not escalate: %v", got)
	}
	st.noteFailureOutcome("ok")
	st.noteFailureOutcome("error") // 窗口内第 3 败（交错分布）
	got := st.escalationReminders()
	if len(got) != 1 || !strings.Contains(got[0], "停止原样重试") {
		t.Fatalf("three failures in the window must escalate: %v", got)
	}
	if again := st.escalationReminders(); len(again) != 0 {
		t.Fatalf("escalation must fire once per run: %v", again)
	}
}

// 检索连续空结果（成功但无实质内容）触发「换关键词或如实写缺失」。
func TestEscalationOnZeroHitSearches(t *testing.T) {
	f := &fakeRunCtx{ctx: context.Background(), prof: core.AgentProfile{Model: core.ModelRoute{Chat: "test-model"}}, execOutput: "未找到相关内容"}
	r := &LLMRunner{Stream: scriptStream(padTurns(
		textDeltaTurn("", toolCall("knowledge_search", `{"query":"a"}`), toolCall("knowledge_search", `{"query":"b"}`)),
		textDeltaTurn("好了。"),
	), nil)}
	if err := r.Execute(f); err != nil {
		t.Fatal(err)
	}
	st := &directRun{l: &LLMRunner{}, inflight: map[string]*toolSlot{}}
	st.noteEvidenceCall("knowledge_search", "ok", "未找到相关内容")
	if got := st.escalationReminders(); len(got) != 0 {
		t.Fatalf("one zero hit must not escalate: %v", got)
	}
	st.noteEvidenceCall("knowledge_search", "ok", "无结果")
	got := st.escalationReminders()
	if len(got) != 1 || !strings.Contains(got[0], "材料中无此信息") {
		t.Fatalf("two zero hits must escalate: %v", got)
	}
	// 有实质结果后计数重置。
	st.noteEvidenceCall("knowledge_search", "ok", "营收 3,200 万元")
	if st.zeroHitSearches != 0 {
		t.Fatal("substantive hit must reset zero-hit count")
	}
}

// ---- E4：结构化 source_id 账本 ----

func TestEvidenceLedgerTagsSource(t *testing.T) {
	f := &fakeRunCtx{ctx: context.Background(), prof: core.AgentProfile{Model: core.ModelRoute{Chat: "test-model"}}, execOutput: "上半年营收 3,200 万元。"}
	r := &LLMRunner{Stream: scriptStream(padTurns(
		textDeltaTurn("", toolCall("knowledge_search", `{"query":"营收"}`)),
		textDeltaTurn("总结完成。"),
	), nil)}
	if err := r.Execute(f); err != nil {
		t.Fatal(err)
	}
	// 结果文本必须带来源标记，且账本登记了对应条目（直接经 runTool 断言）。
	st := &directRun{l: &LLMRunner{}, rc: f, eff: f.Settings(), inflight: map[string]*toolSlot{}}
	res := st.runTool("knowledge_search", `{}`)
	if !strings.Contains(res.Content, "【src_1｜knowledge_search】") {
		t.Fatalf("tool result must carry the source tag: %.120s", res.Content)
	}
	if len(st.ledger) != 1 || st.ledger[0].ID != "src_1" || !strings.Contains(st.ledger[0].Text, "3,200") {
		t.Fatalf("ledger = %+v", st.ledger)
	}
	// 第二次登记递增编号。
	res2 := st.runTool("fs_read_file", `{}`)
	if !strings.Contains(res2.Content, "【src_2｜fs_read_file】") {
		t.Fatalf("second tag must be src_2: %.120s", res2.Content)
	}
}

// 非证据工具不登记、不加标记。
func TestEvidenceLedgerSkipsNonEvidenceTools(t *testing.T) {
	f := &fakeRunCtx{}
	st := &directRun{l: &LLMRunner{}, rc: f, eff: f.Settings(), inflight: map[string]*toolSlot{}}
	res := st.runTool("time_now", `{}`)
	if strings.Contains(res.Content, "src_") || len(st.ledger) != 0 {
		t.Fatalf("non-evidence tool must not enter the ledger: %.80s %+v", res.Content, st.ledger)
	}
}

// 材料读（inputs/）打标；模型自产文件的读与 grep 不打——尾注会污染
// read→edit→build 编辑循环（E4 首轮 eval 实测教训，回归锚点）。
func TestEvidenceLedgerOnlyTagsMaterialReads(t *testing.T) {
	if !isEvidenceRead("read_file", json.RawMessage(`{"path":"inputs/年报.docx.md"}`)) {
		t.Fatal("inputs/ read must count as evidence")
	}
	if isEvidenceRead("read_file", json.RawMessage(`{"path":"方案.md"}`)) {
		t.Fatal("own-workspace read must NOT count as evidence")
	}
	if isEvidenceRead("grep", json.RawMessage(`{"pattern":"营收"}`)) {
		t.Fatal("grep must never count as evidence")
	}
}